Astronauts in low Earth orbit can pull a flight surgeon onto a live video link in seconds, but for Moon and Mars missions, where a radio signal can take more than twenty minutes each way, NASA is evaluating Red Hat's open source AI model, Ramalama, to make medical calls without…
The astronaut physician on a Mars mission may not be a person. It will probably be an AI model running on the spacecraft, asked to make calls no one on Earth can second-guess in time to matter. The people planning that mission have not yet sorted out who is responsible when it is wrong.
That is the structural break inside NASA's current test of Red Hat's Ramalama model as a medical assistant for future Moon and Mars crews. On the International Space Station today, a crew medical issue gets a ground-based flight surgeon on a live video link within seconds. For missions hundreds of thousands of miles farther out, that model collapses. A signal to Mars can take more than twenty minutes each way, so the doctor on call has to live inside the spacecraft, or be software.
NASA's own research on the operational problem frames it in a technical memorandum on communication-delay medical response. The agency is openly building toward crews that diagnose, treat, and triage without synchronous ground support. Outside analysis of mobile deep-space medical systems describes the broader kit those crews would carry: sensors, treatment hardware, and decision aids. The AI medical assistant is the hardest piece, because it is the one that decides.
The case for doing this now, not later, was sharpened by an event 250 miles below the Moon. In January 2026, a Crew-11 astronaut aboard the ISS developed an unexplained medical issue serious enough to cut the mission short and trigger the station's first medical evacuation. Smithsonian Magazine reported that doctors still had not publicly identified the cause months after Crew-11 returned. Even with a real-time link to the crew, ground-based flight surgeons had not been able to crack the case in orbit. Treat the Crew-11 incident as illustrative rather than confirmed: it shows the diagnostic ceiling even when ground help is one radio call away, and it explains why NASA is investing in tools that do not depend on that link.
That is also where the public announcement runs out of answers. Red Hat says Ramalama is being tested as a NASA medical-assistant candidate, but the model's exact training corpus, autonomy tier, and the metrics NASA is using to evaluate it are not yet public. Neither NASA nor Red Hat has published what clinical authority the AI would have, what data it would draw on, or how a crew would override it. The Register's report on the test describes a research collaboration, not a deployed system.
What is missing from the picture is the human back end. On the ISS, when a flight surgeon makes a call, the chain is well-practiced: NASA is the operator, the surgeon is credentialed, the call is logged, and there is a paper trail. If the call goes wrong, the responsibility is layered and knowable. Move that decision to an AI on a spacecraft that cannot phone home, and every part of that chain needs a new answer. Who certifies the medical model the way a medical device is certified today? Who approves the training data, and is that a clinical decision, a procurement decision, or both? Who carries liability if an AI-recommended treatment harms a crew member halfway to Mars? When the AI disagrees with the astronaut, who wins?
These are not hypothetical questions. The same delay that makes ground control useless also makes an after-action lawsuit in Earth orbit useless. The license, the regulator, and the legal forum all have to exist before the spacecraft flies.
NASA's public framing of the Ramalama test leans technical and clinical. The harder story is structural. The agency is rehearsing the moment when an autonomous system holds a clinical decision for a person no one on the ground can reach, and the accountability scaffolding has not been built yet. Watch for whether future NASA solicitations name a certification pathway for medical AI, whether any published Ramalama evaluation specifies the model's decision authority, and whether partner agencies say out loud who carries the legal weight when the AI is the doctor.